In AI circles, there is a story that is frequently shared at conferences or over coffee after official sessions. It involves Garry Kasparov, the chess grandmaster, sitting across from IBM’s Deep Blue in 1997 and losing. Not just losing a game, but losing something more difficult to identify. Afterwards, Kasparov characterized the experience as coming across “a new kind of intelligence, a spirit in the machine.”
What’s interesting is that Deep Blue was, by today’s standards, breathtakingly dumb. It was limited to playing chess. It didn’t know what a chair was. It was unable to carry on a dialogue. But that moment planted a question that never really went away: when does a machine stop performing tasks and start actually thinking?
| Topic | Artificial General Intelligence (AGI) |
|---|---|
| Full Form | Artificial General Intelligence |
| Concept Origin | Theoretical AI research, mid-20th century |
| Key Figures | Dario Amodei (Anthropic), Geoffrey Hinton (independent), Yann LeCun (Meta), Demis Hassabis (DeepMind) |
| Dario Amodei’s Prediction | Early AGI traits possible by 2026 |
| Survey Consensus (2023) | 50% probability of AGI between 2040–2061 |
| Geoffrey Hinton’s Estimate | 5 to 20 years |
| Yann LeCun’s View | Decades away, possibly never in imagined form |
| Current Leading Models | GPT-4 (OpenAI), Gemini (Google), Gato (DeepMind) |
| AGI vs Narrow AI | AGI generalizes; Narrow AI specializes |
| Key Milestone Referenced | Deep Blue defeats Kasparov (1997) |
| Notable Research Events | AI Impacts Survey 2023, NIPS/ICML Expert Surveys |
| Primary Challenge | Reasoning, memory, world-modeling outside training data |
That question is louder now than it’s ever been. AGI — Artificial General Intelligence — is the version of AI that doesn’t need to be told what to do in each situation. It reasons across domains, adapts to new problems without specific training, and in theory, accumulates and applies knowledge the way a curious person does.
Unlike the chatbots and recommendation systems that already run large parts of daily life, AGI would generalize — pulling a concept from biology and using it to solve a problem in finance, the way a human researcher might. Nobody has built it yet. But plenty of people are arguing about when they will.
According to Anthropic’s founder, Dario Amodei, systems with early AGI traits might emerge as early as 2026. Not everyone is at ease with that startling timeline. The window is between five and twenty years, according to Canadian computer scientist Geoffrey Hinton, who worked at Google for decades before leaving due to serious concerns about AI safety.
This range is broad enough to accommodate nearly any result. The fact that both men are regarded as serious rather than sensationalist is noteworthy. In particular, Hinton has gained a reputation for being correct about issues that were initially disregarded.
Yann LeCun at Meta, on the other hand, has a very different perspective. According to LeCun, AGI is still decades away and might not be possible in the way that most people envision. He might be correct. It’s also possible that the definition of AGI keeps shifting just fast enough to stay out of reach, like a mirage that recedes as you approach it.
Demis Hassabis at DeepMind seems to sit somewhere in the middle — optimistic about the possibility of human-like reasoning in AI within a decade, but careful to add that fundamental breakthroughs in understanding intelligence itself are still missing.
The results of surveys conducted among researchers are informative. A 2023 study by AI Impacts involving nearly 2,800 AI researchers put the 50% probability mark for high-level machine intelligence somewhere between 2040 and 2061. According to a similar survey conducted in 2022, that number is 2059.
The gap between Amodei’s 2026 and the survey median of 2040 is enormous — and yet both camps are populated by serious, credentialed people working on the same underlying technology. It’s not careless thinking that causes the uncertainty. It is an accurate representation of how difficult it is to quantify this issue.
The fact that today’s most impressive models—GPT-4, Gemini, systems that can write code, summarize documents, and have coherent conversations across topics—already feel unsettlingly capable contributes to the difficulty of the debate. However, they continue to be limited in a particular technical sense. Outside of their training, they are unable to exercise true initiative.
They have trouble with reasoning chains that call for maintaining a real-time model of the world in memory. Even if a language model is capable of writing a legal brief, it is still unable to recognize, on its own, that it has misinterpreted the client’s circumstances and modify its entire strategy. AGI exists in that gap, which is where existing systems fall short, between performing well and truly knowing what you’re doing.
Observing all of this, it’s difficult to ignore the significance of the experts’ disagreement. There isn’t a communication issue when those closest to the technology can’t agree on a decade, much less a year. It is an indication that something truly unknown is in jeopardy. The Deep Blue moment is beginning to resemble a prologue, a first act whose greater significance is still being worked out.